A deep reinforcement learning approach for early classification of time series

This repository contains an implementation of the paper of Martinez Coralie, Guillaume Perrin, E Ramasso, Michèle Rombaut. A deep reinforcement learning approach for early classification of time series. 26th European Signal Processing Conference (EUSIPCO2018), Sep 2018, Rome, Italy. hal-01825472

Data

The data used in this project is the GunPoints datasets which comes from the UCR/UEA archive

Code

The code is divided as follows:

• The Train.py python file contains the necessary code to run the training step
• The Agent.py python file contains the necessary code about the Agent (used model, memory, ...)
• The Env.py python file contains all code about the environement used for reinforcement learning(reward, state, step, ...)
• The utils.py python file contains all necessary utils function (loss_function and zeros_padding function)
• The Import.py python file contains all necessary package to load

To run the training of model you should issue this following command:

$ python Train.py 

Prerequistes

All python packages needes are listed in pip-requirement.txt file and can be installed simply using the pip command:

• numpy version 1.18.1 or later
• pandas version 1.0.0 or later
• scikit-learn version 0.22.1 or later
• scipy version 1.4.1 or later
• matplotlib version 3.1.3 or later
• pyts version 0.10.0 or later
• tensorflow-gpu version 1.15.0 or later
• keras version 2.2.4 or later
• h5py version 2.10.0 or later